A physics-informed super-resolution network
Date and Time: Monday, May 17, 2021, 03:00pm -
Speaker: Diane Salim (Rutgers University)
We investigate the plausibility of developing a deep-learning-based pipeline to realistically increase the resolution of turbulent fluid simulations of Rayleigh-Bernard convection. We expand upon the method detailed in Jiang et al 2020, which combines learning image-based features from a convolutional neural net (namely, UNet) with information from the underlying partial differential equations of the simulations to strive towards a more physically meaningful super-resolution method. In this work, we expand the range of turbulent scenarios considered, namely by creating a training set with Rayleigh numbers ranging from 10^6 - 10 ^10. We also make modifications to include a more physics-informed architecture by including boundary condition penalization and the solenoidal constraint on the velocity field. We demonstrate the impacts of the including the boundary conditions in the training process. Additionally, we explore the role of the mass continuity equation on the training process and assess its significance as the Ra number increases. In addition to analyzing predictions from the physics-aware network, we also discuss the consequences on network training.